You Have Systems of Record. You Don’t Have a System of Work.
Why AI hasn’t moved the needle yet for most finance teams, and what to do about it.
Every finance team I talk to has tried AI by now. They’ve pasted data into ChatGPT, asked Claude to write a formula, had Gemini summarize a contract. They get a reasonable answer, nod, and go back to the spreadsheet they were working in before. Six months later, nothing has materially changed.
I don’t think this is because AI isn’t powerful enough. I think it’s because we’re pointing it at the wrong part of the workflow.
Watch how work actually gets done
Think about a specific process on your team. Pick the one that makes you wince during close. Maybe it’s bank reconciliation, maybe it’s intercompany, maybe it’s preparing a board deck. Now trace how it actually works. The real version, not the process doc.
Someone pulls a report from NetSuite. Someone else downloads a CSV from the bank portal. A third person digs up a contract from a shared drive. All of this lands in a spreadsheet. That spreadsheet becomes the workbench.
Your team normalizes the data, stitches it together, applies business logic that often lives in someone’s head. They build formulas to get to the number they need. Then they send the result somewhere downstream. Post a journal entry, update a report, send a collections email, flag something for review.
The ERP, the bank, the CRM. Those are your systems of record. They store data. They’re essential. But they are not where the thinking happens.
The thinking happens in the gaps between those systems. In the spreadsheet where someone is contextualizing data from three different sources. In the email thread where someone is chasing down a missing remittance. In the manual process that takes two days every month because nobody has built a real system around it.
You have systems of record. What you don’t have is a system of work.
Where AI in finance actually stands
ERP vendors are making real investments in AI, and some of what they’re shipping is useful.
Chatbots on top of your data. The best implementations use curated semantic layers, validated queries, governed data models. You can get quick answers to ad hoc questions, and that’s a real time saver. But even the best chatbot has a ceiling. For the numbers you care about most, you want a reliable dashboard with drill-down into underlying transactions, not a chat answer you hope is right. Chat is a great interface for exploration. It’s not a great interface for running your close.
AI-enabled processes inside existing systems. Revenue contract review inside an ERP, automated transaction coding, anomaly detection. When your business logic fits neatly into what the vendor has built, this works. But I’ve watched teams get excited about an AI feature, test it against their real data, realize it doesn’t handle their edge cases, and go right back to doing things manually. Complex business logic is the norm, not the exception, and vendor implementations can only cover so much ground before they hit the wall of your specific business.
Connectors and integrations. Your ERP gives you an MCP connector, so now you can pull NetSuite data into a chatbot. This is genuinely cool infrastructure. But having access to the data is table stakes. The question is what you do with it, and a chat window isn’t where your team is going to manage AR aging or build a close deck.
These tools are doing real things. But they all share the same limitation: they’re adding AI to systems of record. And the hardest, most time-consuming work your team does isn’t inside any single system of record. It’s in the space between all of them.
The trust problem
If you can’t explain it to your auditor, it’s not production-ready. That’s the bar in finance, and it’s the part of the AI conversation that gets glossed over the most.
When a controller evaluates any new tool, the first question isn’t “is this cool?” It’s “can I trust the output enough to put my name on it?” If a model is generating a number, how do I validate it? If an agent is coding a transaction to a GL account, what’s the audit trail? If something goes wrong, can I trace back to the decision point and understand why?
Most AI tools available today don’t answer these questions well. And that’s why smart finance teams try them, see the potential, and then quietly go back to the way things were. It’s not resistance to change. It’s a rational response to insufficient controls.
Any serious AI deployment in finance needs full audit trails, human-in-the-loop approvals for high-stakes actions, and the ability for someone other than the original builder to understand what’s happening and why.
What to actually do about it
If you’re a controller, a VP of Finance, or a CFO who’s seen the promise of AI but hasn’t felt the impact, here’s where I’d start.
Pick one workflow that’s painful. A specific process, not a vague statement like “we want to automate month-end close.” Bank reconciliation. Vendor invoice triage. Collections follow-up. Month-end flux analysis. Something your team does repeatedly that eats time.
Then trace the real workflow end to end. Four questions:
What data comes in, and from where? List every source, including the unstructured stuff. The ERP report, the bank CSV, the email attachment, the Slack message explaining why an invoice is higher than expected. The unstructured context is usually the most important.
Who stitches this together, and how? It’s almost always a spreadsheet. Someone normalizing data from different sources, cross-referencing, building the picture that no single system provides.
What reasoning and judgment gets applied? The rules that determine how a transaction should be coded, whether an invoice needs escalation, what the right dunning cadence is for a given customer.
Where does the output go? A journal entry posted to the ERP. An email sent to a customer. A dashboard updated for leadership. A file uploaded for audit.
When you map this out for even one process, you’ll see it clearly. The hardest work is happening in the space between your systems, in spreadsheets and email and institutional memory. That’s where there’s no real system today. And that’s exactly where AI can do the most.
What this looks like when it works well
One team on Lumera was spending two days per month on bank reconciliation. Download bank statements, pull transaction data from NetSuite, open a spreadsheet, match deposits against supporting docs, figure out the right GL accounts and departments based on business rules the team had memorized, then manually key journal entries into the ERP. They built an AI workflow that ingests bank data and supporting documents, applies their specific coding logic, surfaces exceptions for human review, and exports clean journal entries ready for posting. The matching and coding that took two days now happens in minutes. The accountant’s job shifted from data entry to review and approval, which is what it should have been all along.
Another team wanted a custom dashboard on Netsuite data with full drill-down into underlying transactions, built around how they think about cash, not how their ERP thinks about cash. They coded a fully custom live dashboard on Lumera, connected directly to their NetSuite data, with the exact views and drill-downs they needed. No more exporting to Excel every week to build the view manually.
A third team built an AI collections agent that looks at AR aging, applies their dunning logic (gentle reminder vs. firmer follow-up vs. escalation), considers payment history and relationship context, and drafts emails for review. Hours of composing collection emails turned into minutes of reviewing and sending.
In each case, the team started by mapping the real workflow. They identified where context was being assembled manually, where judgment was being applied, and where the output needed to go. Then they built a system of work around that, with the audit trails, approvals, and controls that production finance requires.
The opportunity
The gap between systems has always existed because every company’s business is different. The long tail of finance work, the processes that don’t fit neatly into any single product, is enormous. No platform can anticipate every chart of accounts structure, every allocation methodology, every close process quirk. That’s not a failure of the tools. It’s the nature of the work.
AI is the first technology that can actually operate in that long tail. It can parse unstructured data, apply nuanced business logic, pull context from multiple systems, and generate outputs in the right format for wherever they need to go. But only if you point it at the right problem. Not “add a chatbot to the ERP.” Instead: build a real system of work for the cross-system, high-judgment processes where your team is spending their time today.
Your systems of record aren’t going anywhere. The system of work you’ve never had is what’s ready to be rebuilt.
Sowmya is the CEO and co-founder of Lumera, AI infrastructure for finance teams. She was previously Controller at OpenAI and Rippling, and led Corporate Accounting at Square.